雾无线接入网中基于神经网络的资源分配方案

考虑雾无线接入网（Fog Radio Access Network,F-RAN）中的性能优化问题,提出一种基于深度神经网络（Deep Neural Network,DNN）的资源分配方案。该方案旨在通过资源分配策略来最大化经济频谱效率（Economical Spectral Efficiency,ESE）。为解决传统资源分配方案需要大量计算的问题,该方案借助神经网络模型,将ESE作为损失函数,使用更少的计算量来确定用户的波束赋形,从而实现实时处理。仿真结果表明,相比于基于传统凸优化功率分配方案或者是基于监督学习的CNN方法,所提出的方案的光谱效率（Spectral Efficiency,SE）和ESE的最大增益分别可以达到5%和20%。此外,该方案在执行时间上与CNN方案接近,明显优于传统算法。     

结合局部全局特征与多尺度交互的三维多器官分割网络

目的 高度适形放射治疗是常用的癌症治疗方法,该方法的有效性依赖于对癌组织和周边多个危及器官（organ at risk,OAR）解剖结构的精确刻画,因此研究三维图像多器官的高精度自动分割具有重要意义。以视觉Transformer(vision Transformer,ViT)和卷积神经网络（convolutional neural network,CNN）结合为代表的三维医学图像分割方法表现出了丰富的应用优势。然而,这类方法往往忽略同一尺度内和不同尺度间的信息交互,使得CNN和ViT特征的提取和融合受限。本文提出一种端到端多器官分割网络LoGoFUNet(local-global-features fusion UNet),旨在应对现有方法的缺陷。方法 首先,针对单一器官分割,提出在同一尺度下并行提取并融合CNN和ViT特征的LoGoF(local-global-features fusion)编码器,并构建了一个端到端的三维医学图像分割多尺度网络M0。此外,考虑到器官内部以及器官之间的相互关系,该方法在M0网络的基础上设计并引入了多尺度交互（multi-scale interacti...     

Systems Approach to Pathogenic Mechanism of Type 2 Diabetes and Drug Discovery Design Based on Deep Learning and Drug Design Specifications

In this study, we proposed a systems biology approach to investigate the pathogenic mechanism for identifying significant biomarkers as drug targets and a systematic drug discovery strategy to design a potential multiple-molecule targeting drug for type 2 diabetes (T2D) treatment. We first integrated databases to construct the genome-wide genetic and epigenetic networks (GWGENs), which consist of protein–protein interaction networks (PPINs) and gene regulatory networks (GRNs) for T2D and non-T2D (health), respectively. Second, the relevant “real GWGENs” are identified by system identification and system order detection methods performed on the T2D and non-T2D RNA-seq data. To simplify network analysis, principal network projection (PNP) was thereby exploited to extract core GWGENs from real GWGENs. Then, with the help of KEGG pathway annotation, core signaling pathways were constructed to identify significant biomarkers. Furthermore, in order to discover potential drugs for the selected pathogenic biomarkers (i.e., drug targets) from the core signaling pathways, not only did we train a deep neural network (DNN)-based drug–target interaction (DTI) model to predict candidate drug’s binding with the identified biomarkers but also considered a set of design specifications, including drug regulation ability, toxicity, sensitivity, and side effects to sieve out promising drugs suitable for T2D.     

Integration of deep neural networks and ensemble learning machines for missing well logs estimation

Geophysical logging is one of the most important measurement techniques for the oil/gas development and exploration industry. In practice, missing well logs estimation/prediction or soft logging is one of the effective ways to save oil/gas exploration costs. Due to the structural complexity and heterogeneous of the geological reservoir, there must be strong nonlinear relationships among different well logs. In order to reveal one or more of these relationships, multiple linear regressions, Bayesian learning and traditional machine learning methods (ANN, SVM, etc.) are always employed in the literature. However, in practice, it is impossible to obtain a compact data set that accurately reflects these nonlinear relationships. Therefore, falling into local optimum solution is the fatal defect of these traditional methods. In order to address this problem for a certain extent, we propose to integrate deep neural networks (DNN) and several ensemble learning machines (ELM) to reveal these relationships more accurately. Experiential results illustrated that the proposed method can really estimate missing logs more accurately than traditional ones, and the performance is promising.     

基于SDAE和双模型联合训练的低压用户窃电检测方法

用户窃电行为是电网企业运营管理的痛点,基于数据驱动的低压用户窃电检测是当前的重要发展方向.由于窃电数据集具有自身高维度且样本不平衡的特点,对窃电检测模型的拟合能力和泛化能力要求极高.为此,文章利用堆栈降噪自编码器对低压用户日用电量数据进行特征提取,通过挖掘数据的深层特征减少窃电产生的极端数据对检测模型的影响;进而提出逻辑回归与深度神经网络联合训练模型进行低压用户窃电检测,将逻辑回归模型的记忆能力与深度神经网络模型的泛化能力相结合,进一步提升窃电检测的精度.通过实际电网数据的实验仿真,从AUC值、准确率和召回值三个评价指标验证了所提出方法相对于传统机器学习算法具有明显的性能优势.     

GSO:基于图神经网络的深度学习计算图子图替换优化框架

深度学习在各种实际应用中取得了巨大成功,如何有效提高各种复杂的深度学习模型在硬件设备上的执行效率是该领域重要的研究内容之一.深度学习框架通常将深度学习模型表达为由基础算子构成的计算图,为了提高计算图的执行效率,传统的深度学习系统通常基于一些专家设计的子图替换规则,采用启发式搜索算法来优化计算图.它们的不足主要有:1)搜...     

计及N-1安全约束的输电网阻塞管控数据驱动模型

负荷的快速增长及其时空分布的快速变化导致输电阻塞频繁发生,影响电力系统安全运行。在传统的阻塞管理模型中,由于N-1约束的强非凸非线性,对于大型系统耗时较长,难以应对快速变化的运行工况。因此提出采用深度网络拟合N-1约束,嵌入到优化控制模型中,提高求解效率。基于佳点集理论产生控制样本,以发电机控制量及负荷水平作为输入,以N-1违约量作为输出,基于深度神经元网络(Deep Neural Network,DNN)构建N-1评估器,嵌入到传统阻塞管理模型中,并采用遗传算法求解该模型,以IEEE30节点为算例验证方法的有效性,仿真结果表明该方法较传统方法而言,有效提高了收敛时间,同时又保证了传统模型的求解精度。     

Adaptive fast and targeted adversarial attack for speech recognition

Adversarial examples are malicious inputs designed to induce deep learning models to produce erroneous outputs,which make humans imperceptible by adding small perturbations to the input.Most research on adversarial examples is in the domain of image.Recently,studies on adversarial examples have expanded into the automatic speech recognition domain.The state-of-art attack on the ASR system comes from C &W,which aims to obtain the minimum perturbation that causes the model to be misclassified.However,this method is inefficient since it requires the optimization of two terms of loss functions at the same time,and usually requires thousands of iterations.In this paper,we propose an efficient approach based on strategies that maximize the likelihood of adversarial examples and target categories.A large number of experiments show that our attack achieves better results with fewer iterations.     

基于神经网络平均影响值的超短期风电功率预测

针对动态神经网络风电功率预测模型输入变量较多、模型复杂的问题,将神经网络和平均影响值方法相结合,提出了一种基于神经网络平均影响值的超短期风电功率预测方法。此方法综合考虑了各输入变量对输出变量(风电预测功率)的外部贡献率和内部贡献率,筛选出了对输出变量贡献率最大的输入变量,建立了一个优化的神经网络超短期风电功率预测模型。实验结果表明,所提模型降低了预测模型的复杂度,减少了测量噪声对预测精度的影响,得到了较好的风电功率预测结果。